Why not just use SELECT with a bounding box and an index on lat and lon?

A composite index scans on its leading column, so the engine seeks the latitude band and then row-filters longitude. Near London a 5 km box over 8 million rows selects around 40,000 rows to return 1,000. It is correct and it is fast enough until it is not, and the failure arrives as steadily rising p99 rather than as an error.

How stale is the search index allowed to be?

A few seconds, and you should say so explicitly. A new listing appearing 3 seconds late has no user-visible consequence at a 20,000 to 1 read-write ratio. Deletes are the exception: a closed venue in results generates real complaints, so route deletions as high-priority events rather than letting them wait for a rebuild.

Does this design work for Uber?

No, and saying why is a good answer. Driver positions change every few seconds, so a disk-backed index rebuilt asynchronously is the wrong structure. Ride-hailing keeps current locations in memory, updates them continuously, and uses the geospatial cell only as a bucket key. The indexing idea carries over; the storage and freshness assumptions do not.

Common Designschapter 8 of 8 · 4 lessons

Proximity Service (Yelp, Nearby)

Find businesses within a radius, ranked by distance, over hundreds of millions of rows.

← All chapters View in cheatsheet

Proximity search is the interview question that looks small and is not. Find restaurants within 5 km, nearest first.

The geospatial index is the hard part everybody expects. What separates a junior answer from a senior one is everything around it: how the data gets split across machines, how it stays fresh, and what happens in Manhattan.

Lessons

4 in this chapter

Sizing the ProblemEstimate first, because the numbers rule out most of the designs.2 min
The Read PathTap to ranked list: nine cell lookups, one distance pass, one sort.2 min
Splitting Reads from WritesA 20,000 to 1 ratio means the two paths should not share a table.2 min
Density and HotspotsPopulation is not uniform, so a uniform grid shards badly.3 min